Recent Developments in Ultrasound-based Bone Registration

June 21-24, 2006, Montreal, Canada INTRODUCTION: Ultrasound (US) imaging has been proposed as a non-invasive method for bone registration during CAOS [1,2,3]. In this technique, an US probe tracked by a 3D localiser acts as a non-invasive bone localisation device. Since the bone surface does not need to be surgically exposed, and a much larger surface area may be sampled, this method is potentially more accurate and faster than conventional direct contact techniques. In most conventional approaches to US-based registration, an A- or B-mode US probe is used to digitise bone surface points [1,3]. In practice, manual identification is still the most reliable way of identifying points from US data, but is too time-consuming to be clinically practical, and US image artefacts make accurate and reliable bone surface identification a challenging task to automate. Furthermore, variations in acoustic properties of soft tissue overlying bone, such as the speed of sound, can introduce significant errors (>5%), which may be particularly problematic in the obese patient where they can translate into a bone localisation error on the order of several millimetres. In this study, we validated two new innovations that address the problems outlined above. In the first, in vivo localisation errors are minimised by adopting a self-calibrating registration algorithm. In the second, the registration algorithm is fully automatic. Full details of the experimental set-up and registration algorithms can be found in [4,5]. METHODS: Three intact human cadavers were used to validate the accuracy of US-based registration of the femur and pelvis to a preoperative CT scan. Between 248 and 576 tracked B-mode US images were acquired for each of the 6 femurs and 3 pelves using an experimental 3D US system developed in our research group. Bone-implanted dynamic reference objects (DROs) were used to define a physical 3D co-ordinate system for each bone. The US probe, DRO and digitizer were all tracked using an NDI Optotrak 3020 localiser. The rigid-body CT-to-patient transformation was determined in three ways: Firstly, by paired-point matching bone-implanted fiducial markers, identified in CT and in physical co-ordinates. The result was used as a Gold Standard transformation to evaluate the accuracy of two US-based registration methods. Secondly, registrations were performed using a self-calibrating version of a standard point-to-surface registration algorithm [4]. A novel feature of this algorithm is the ability to update calibration parameters of the 3D US system. The algorithm was run in two modes: in the first, only the pixel scaling parameter in the direction of the US beam (which is directly related to the average speed of sound) was updated, whereas in the second all calibration parameters were updated. Initial estimates of the calibration parameters were obtained using a phantom. To ensure accurate identification of the bone surface, manual segmentation was used to extract surface points from the US images and a surface from the CT scan. Lastly, registrations were performed using a fully automatic algorithm, which, once trained using independent data, transforms the US and CT images into images where each grey-level pixel/voxel value represents the probability of a bone surface at that location [5]. The algorithm was also self-calibrating, but limited to updating the US axial scaling parameter. For each of the 9 bones, 100 registrations were performed using random initial estimates of the rigid-body parameters determined by simulating paired-point matching of anatomical skin landmarks. The registration accuracy for each bone was calculated as the average root-mean-square target registration error (TRE), computed over the whole bone surface, from 100 trials. At any point on the surface, the TRE may be interpreted as the distance between the point position after registering using the Gold Standard transformation and the point position after registering using the US-based method. The estimated mean TRE of the Gold Standard registration method was 0.8mm. RESULTS: Using the US-point-based registration algorithm, the mean (min-max) TRE over all 9 bones was 2.4 (1.1–3.8) and 2.0 (1.0–3.1) mm, with and without updating the axial scaling parameter, respectively. The corresponding errors for the automatic algorithm were 1.8 (1.0–3.1) and 1.6 (0.7–3.0) mm. Compared with the standard method, updating the scaling parameter during the registration led to a reduction in TRE of up to 33% in the case of the point-based algorithm, and up to 45% for the automatic algorithm. Updating all calibration parameters led to further improvements in the accuracy of the point-based method, yielding a mean TRE of 1.6 (0.5–2.8) mm. Between 42 and 52 of the 900 point-based registrations failed, whereas just one automatic registration failed. CONCLUSIONS: Compared with a conventional US-point-based algorithm, the automatic registration algorithm evaluated in this study was found to be more robust and have slightly better accuracy. Since, this method requires no user interaction following US scanning, it is much more useful in the clinical situation, although the processing time (currently 2–10 mins) needs to be reduced to a clinically acceptable level. Self-calibration improved the accuracy of both methods, in certain cases leading to a large reduction in TRE. We believe that both of these developments constitute significant progress towards making US-based registration a clinically practical tool for computer-assisted procedures. Future work will focus on developing these techniques to achieve this objective. REFERENCES 1. S. Lavallee, et al., "Computer-assisted spinal surgery using anatomy-based registration," in Computer Integrated Surgery: Technology and Clinical Applications, Taylor RH, Lavallee S, Burdea GC, and Mosges R (Eds.), Cambridge, MA: MIT Press, pp. 425-446, 1996. 2. D. Amin, et al., "Ultrasound registration of the bone surface for surgical navigation," Comp. Aid. Surg., vol. 8, pp. 1-16, 2003. 3. S. Heger, et al., "User-Interactive Registration of Bone with A-mode Ultrasound," IEEE Eng. Med. Biol. Mag., vol. 24, pp. 85-95, 2005. 4. D.C. Barratt, et al. "Self-calibrating 3D-ultrasound-based bone registration for minimally invasive orthopedic surgery". IEEE Trans. Med Imag., vol. 25, No. 3, 2006 (In Press) and Proc. MICCAI 2005, LNCS 3749, Springer, pp. 605-612, 2005. 5. G.P. Penney, et al., "Cadaver validation of intensity-based ultrasound-to-CT registration," Med. Imag. Anal., 2006 (In Press) and Proc. MICCAI 2005, LNCS 3750, Springer, pp. 1000-1007, 2005.